EP2416254A1 - Arbitration method, arbitration circuit, and device using the arbitration circuit - Google Patents

Arbitration method, arbitration circuit, and device using the arbitration circuit Download PDF

Info

Publication number
EP2416254A1
EP2416254A1 EP09842617A EP09842617A EP2416254A1 EP 2416254 A1 EP2416254 A1 EP 2416254A1 EP 09842617 A EP09842617 A EP 09842617A EP 09842617 A EP09842617 A EP 09842617A EP 2416254 A1 EP2416254 A1 EP 2416254A1
Authority
EP
European Patent Office
Prior art keywords
arbitration
algorithm
arbitrary
plural
input ports
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP09842617A
Other languages
German (de)
French (fr)
Other versions
EP2416254A4 (en
Inventor
Makoto Hataida
Toshikazu Ueki
Takayuki Kinoshita
Yoshikazu Iwami
Hidekazu Osano
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fujitsu Ltd
Original Assignee
Fujitsu Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fujitsu Ltd filed Critical Fujitsu Ltd
Publication of EP2416254A1 publication Critical patent/EP2416254A1/en
Publication of EP2416254A4 publication Critical patent/EP2416254A4/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L49/00Packet switching elements
    • H04L49/10Packet switching elements characterised by the switching fabric construction
    • H04L49/101Packet switching elements characterised by the switching fabric construction using crossbar or matrix
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L49/00Packet switching elements
    • H04L49/25Routing or path finding in a switch fabric
    • H04L49/253Routing or path finding in a switch fabric using establishment or release of connections between ports

Definitions

  • the arbiter circuit 18 described above may appear to perform the arbitration equally, however, when one focuses on a certain packet, a queue may not be output for a long time from the queue arbiter circuit 18-1 to the inter-port arbiter circuit 18-2, to thereby generate the so-called livelock.
  • the livelock may be prevented, but the arbitration time becomes long to deteriorate the performance of the parallel computer 1.
  • an apparatus to couple a system having a plurality of processors and memories to an input output system having a plurality of input output interfaces includes physically plural input ports logically having plural virtual channels; plural output ports; a first arbiter circuit configured to perform a path control to transfer data from the plural input ports to an arbitrary output port amongst the plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm; a second arbiter circuit configured to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and an arbitration switching circuit configured to switch the arbitrary arbitration algorithm used in the first and second arbiter circuits to the time-division algorithm for a predetermined time in response to a trigger.
  • the increase in the arbitration time may be suppressed and the livelock may be prevented.
  • This embodiment differs from the first embodiment described above in that, empty signals EMPTY_A, EMPTY_B, EMPTY_C and EMPTY_D respectively indicating that the queues AQ, BQ, CQ and DQ at each of the input ports AI, BI, CI and DI are empty are supplied from the buffer to an arbitration switching circuit 30B.
  • an inhibit signal INH that inhibits input or insertion of the packet is supplied to each of the input ports AI, BI, CI and DI.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)
  • Multi Processors (AREA)

Abstract

An arbitration method includes a first arbitration process to perform a path control to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary one of the plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm, and a second arbitration process to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm. The arbitrary arbitration algorithm used in the first and second arbitration processes is switched to the time-division algorithm for a predetermined time in response to a trigger.

Description

    TECHNICAL FIELD
  • The present invention generally relates to arbitration methods, arbiter circuits, and apparatuses provided with arbiter circuits, and more particularly to an arbitration method, an arbiter circuit, and an apparatus provided with an arbiter circuit, that are suited for a parallel computer or the like.
  • BACKGROUND ART
  • FIG. 1 is a block diagram illustrating an example of a parallel computer. A parallel computer 1 includes a plurality of System Boards (SBs) 12, a crossbar switch 14, and a plurality of Input Output Boards (IOBs) 15 that are connected as illustrated in FIG. 1. Each SB 12 includes a plurality of Central Processing Units (CPUs) 11, and a plurality of memories 13. Each IOB 14 includes a plurality of input parts and a plurality of output parts (or input output interfaces). The crossbar switch 14 includes an input port AI and an output port AO connected to one SB 12, an input port BI and an output port BO connected to the other SB 12, an input port CI and an output port CO connected to one IOB 15, and an input port DI and an output port DO connected to the other IOB 15.
  • The crossbar switch 14 includes the plurality of input ports and the plurality of output ports described above, and performs a routing (path control) in order to transfer packets as data from an arbitrary node, such as the SB 12 and the IOB 15, to another node. In order to avoid a deadlock, the crossbar switch 14 has a plurality of virtual channels for each port. In other words, the crossbar switch 14 physically includes a plurality of ports, and each port logically includes a plurality of channels (that is, virtual channels), but only one channel may be selected at one port at an arbitrary point in time.
  • An arbiter circuit in the crossbar switch 14, that performs an arbitration on the packets from the plurality of ports and the plurality of channels, ideally treats all ports and all channels equally. FIG. 2 is a diagram illustrating such an ideal arbiter circuit. In FIG. 2, an arbiter circuit 17 performs an arbitration process with respect to inputs of channels C0 and C1 from an input port AI, inputs of channels C0 and C1 from an input port BI, inputs of channels C0 and C1 from an input port CI, and inputs of channels C0 and C1 from an input port DI, and outputs a routing request (path control request) from one input port to an arbitrary output port based on a result of the arbitration process.
  • However, in actual circuit design, it is physically difficult to create the arbiter circuit 17 having the structure illustrated in FIG. 2, and it is also difficult to take into consideration the signal delay in the design. For this reason, an arbiter circuit having a 2-stage structure illustrated in FIG. 3 has been proposed.
  • FIG. 3 is a diagram illustrating an example of an arbiter circuit. In FIG. 3, queue arbiter circuits 18-1 select packets from queues AQ, BQ, CQ and DQ for each of the input ports AI, BI, CI and DI, and inter-port arbiter circuits 18-2 select one port from the plurality of input ports AI, BI, CI and DI. The queues AQ, BQ, CQ and DQ are retained in corresponding buffers (not illustrated) within the crossbar switch 14, and blocks identifying the queues AQ, BQ, CQ and DQ in FIG. 3 correspond to these buffers. Hence, an arbiter circuit 18 has the 2-stage structure formed by two kinds of arbiter circuits 18-1 and 18-2.
  • FIG. 4 is a diagram illustrating a structure of the queue arbiter circuit 18-1. As illustrated in FIG. 4, an arbitration algorithm equally arbitrates a set of queues for each of the channels C0 and C1. FIG. 4 illustrates a case in which the Least Recently Used (LRU) algorithm is used as the arbitration algorithm. Hence, a selector 181 selectively outputs one of the queues from the channels C0 and C1, depending on an operation result that is obtained by a LRU algorithm part 180 based on the LRU algorithm.
  • The illustration of a structure of the inter-port arbiter circuit 18-2 will be omitted, because the same arbitration algorithm as the queue arbiter circuit 18-1, such as the LRU algorithm, may be used for the arbitration with respect to a set of queues for each of the input ports AI, BI, CI and DI, in place of the set of queues for each of the channels C0 and C1. In a case in which the path control request (routing request) from the queue arbiter circuit 18-1 is not accepted for a long time and a stall state continues, due to insufficient resources and the like, a retry control is performed to once cancel the request and issue another request in order to prevent deadlock.
  • The arbiter circuit 18 described above may appear to perform the arbitration equally, however, when one focuses on a certain packet, a queue may not be output for a long time from the queue arbiter circuit 18-1 to the inter-port arbiter circuit 18-2, to thereby generate the so-called livelock. However, when a time-division algorithm is used for the arbitration algorithm in place of the LRU algorithm, the livelock may be prevented, but the arbitration time becomes long to deteriorate the performance of the parallel computer 1.
    • Patent Document 1: Japanese Laid-Open Patent Publication No. 2001-22711
    DISCLOSURE OF THE INVENTION PROBLEM TO BE SOLVED BY THE INVENTION
  • In the conventional arbitration method, the arbitration time becomes long when an attempt is made to prevent the livelock, and as a result, the performance of the parallel computer may deteriorate.
  • Accordingly, one object of the present invention is to provide an arbitration method, an arbiter circuit, and an apparatus provided with the arbiter circuit, in which the increase of the arbitration time may be suppressed and the livelock may be prevented.
  • MEANS OF SOLVING THE PROBLEM
  • According to one aspect of the present invention, an arbitration method includes a first arbitration process to perform a path control to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary output port amongst plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm; a second arbitration process to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and a switching process to switch the arbitrary arbitration algorithm used in the first and second arbitration processes to the time-division algorithm for a predetermined time in response to a trigger.
  • According to one aspect of the present invention, an arbiter circuit includes a first arbiter circuit configured to perform a path control to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary output port amongst plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm; a second arbiter circuit configured to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and an arbitration switching circuit configured to switch the arbitrary arbitration algorithm used in the first and second arbiter circuits to the time-division algorithm for a predetermined time in response to a trigger.
  • According to one aspect of the present invention, an apparatus to couple a system having a plurality of processors and memories to an input output system having a plurality of input output interfaces, includes physically plural input ports logically having plural virtual channels; plural output ports; a first arbiter circuit configured to perform a path control to transfer data from the plural input ports to an arbitrary output port amongst the plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm; a second arbiter circuit configured to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and an arbitration switching circuit configured to switch the arbitrary arbitration algorithm used in the first and second arbiter circuits to the time-division algorithm for a predetermined time in response to a trigger.
  • BRIEF DESCRIPTION OF THE DRAWING
    • FIG. 1 is a block diagram illustrating an example of a parallel computer;
    • FIG. 2 is a diagram illustrating an ideal arbiter circuit;
    • FIG. 3 is a diagram illustrating an example of an arbiter circuit;
    • FIG. 4 is a diagram illustrating a structure of the queue arbiter circuit;
    • FIG. 5 is a diagram illustrating an example of an arbiter circuit in a first embodiment;
    • FIG. 6 is a diagram illustrating a structure of an arbitration switching circuit;
    • FIG. 7 is a diagram illustrating a structure of a queue arbiter circuit;
    • FIG. 8 is a diagram illustrating an example of a state of a queue AQ at an input port AI;
    • FIG. 9 is a diagram illustrating an arbitration result for each clock based on the LRU algorithm;
    • FIG. 10 is a diagram illustrating an arbitration result for each clock based on the time-division algorithm;
    • FIG. 11 is a diagram illustrating an example of an arbiter circuit in a second embodiment;
    • FIG. 12 is a diagram illustrating a structure of a queue arbiter circuit;
    • FIG. 13 is a diagram illustrating a structure of an arbitration switching circuit;
    • FIG. 14 is a diagram illustrating an example of an arbiter circuit in a third embodiment; and
    • FIG. 15 is a diagram illustrating a structure of an arbitration switching circuit.
    DESCRIPTION OF REFERENCE NUMERALS
  • 1
    Parallel Computer
    11
    CPU
    12
    SB
    13
    Memory
    14
    Crossbar Switch
    15
    ICB
    28, 28A, 28B
    Arbiter Circuit
    28-1, 28-1A
    Queue Arbiter Circuit
    28-2
    Port Arbiter Circuit
    30, 30A, 30B
    Arbitration Switching Circuit
    BEST MODE OF CARRYING OUT THE INVENTION
  • According to the disclosed arbitration method, arbiter circuit, and apparatus provided with the arbiter circuit, when performing a routing (path control) in order to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary output port amongst plural output ports, only one channel is selectable at one input port at an arbitrary point in time. A first arbitration process performs the arbitration of the channels of each of the input ports according to an arbitrary arbitration algorithm other than the time-division algorithm, and a second arbitration process performs the arbitration of the input ports according to the arbitrary arbitration algorithm.
    In addition, the arbitration algorithm used in the first and second arbitration processes is switched to the time-division algorithm for a predetermined time in response to a trigger.
  • Hence, the increase in the arbitration time may be suppressed and the livelock may be prevented.
  • A description will now be given of embodiments of the disclosed arbitration method, arbiter circuit, and apparatus provided with the arbiter circuit, by referring to FIG. 5 and the subsequent figures.
  • EMBODIMENTS
  • FIG. 5 is a diagram illustrating an example of an arbiter circuit in a first embodiment. It is assumed that an arbiter circuit 28 illustrated in FIG. 5 is provided within the crossbar switch 14 of the parallel computer 1 illustrated in FIG. 1, for example. In FIG. 5, those parts that are the same as those corresponding parts in FIG. 1 are designated by the same reference numerals, and a description thereof will be omitted.
  • The arbiter circuit 28 illustrated in FIG. 5 includes queue arbiter circuits 28-1, port arbiter circuits 28-2, and an arbitration switching circuit 30. The queue arbiter circuits 28-1 select packets from queues AQ, BQ, CQ and DQ for each of the input ports AI, BI, CI and DI, and the port arbiter circuits 28-2 select one port from the plurality of input ports AI, BI, CI and DI. Hence, the arbiter circuit 28 has the 2-stage structure formed by two kinds of arbiter circuits 28-1 and 28-2.
  • The queue arbiter circuit 28-1 performs a queue arbitration based on a first arbitration algorithm or a second arbitration algorithm instructed from the arbitration switching circuit 30. Similarly, the port arbiter circuit 28-2 performs a port arbitration based on the first arbitration algorithm or the second arbitration algorithm instructed from the arbitration switching circuit 30.
  • When an arbitration request from the queue arbiter circuit 28-1 to the port arbiter circuit 28-2 is not accepted (or selected) for a predetermined time, the queue arbiter circuit 28-1 once cancels this arbitration request and reissues another arbitration request. The cancellation of the arbitration request is notified by a retry indicated by dotted lines in FIG. 5.
  • The arbitration switching circuit 30 monitors a frequency of the retry at each of the input ports AI, BI, CI and DI, and outputs the instruction to switch the arbitration algorithm of the arbiter circuits 28-1 and 28-2 from the first arbitration algorithm to the second arbitration algorithm when the frequency of the retry reaches a threshold value. The arbitration switching circuit 30 supplies an arbitration switching signal ARB_EN and a selection signal ARB_SEL1 instructing the arbitration algorithm to the queue arbiter circuit 28-1, and supplies the arbitration switching signal ARB_EN and a selection signal ARB_SEL2 instructing the arbitration algorithm to the port arbiter circuit 28-2. When the retry frequency reaches the threshold value at each of the input ports AI, BI, CI and DI, the arbitration switching circuit 30 asserts the arbitration switching signal ARB_EN that enables the switching of the arbitration algorithm, in order to instruct the switching of the arbitration algorithms of the arbiter circuits 28-1 and 28-2 to the arbitration algorithms instructed by the selection signals ARB_SEL1 and ARB_SEL2.
  • The first and second arbitration algorithms are mutually different arbitration algorithms, and in this embodiment, the first arbitration algorithm is the LRU algorithm, and the second arbitration algorithm is the time-division algorithm. Of course, the first arbitration algorithm is not limited to the LRU algorithm, and for example, it is possible to use other algorithms such as a round robin algorithm that makes the selection according to round robin, a fixed priority algorithm that makes the selection according to a fixed priority order, and the like.
  • FIG. 6 is a diagram illustrating a structure of the arbitration switching circuit 30. The arbitration switching circuit 30 includes a free-run counter 31, a frequency monitoring counter 32, OR circuits 33 and 34, a counter 35 to measure a predetermined time, and a decoder 36 that are connected as illustrated in FIG. 6. In this example, the counters 31, 32, and 35 may be formed by a 10-bit counter, a 4-bit counter, and a 10-bit counter, respectively.
  • When a counted value of the free-run counter 31 reaches a maximum value (for example, 1024 cycles), a clear signal from the free-run counter 31 is input to a clear terminal clr of the frequency monitoring counter 32 to clear the frequency monitoring counter 32. The retry instruction with respect to each of the input ports AI, BI, CI and DI is input to a count-up terminal up of the frequency monitoring counter 32 via the OR circuit 33, to count by the frequency monitoring counter 32 the number of retry instructions with respect to each of the input ports AI, BI, CI and DI. When the counted number of retry instructions or the retry frequency reaches a threshold value (for example, 15 times/1024 cycles), the frequency monitoring counter 32 inputs a signal to a count-up terminal up of the counter 35 via the OR circuit 34, to start the count of the counter 35. The counter 35 asserts the arbitration switching signal ARB_EN output therefrom while the counter 35 counts the predetermined time (that is, until a counted value exceeds a predetermined value). During the time in which the arbitration switching signal ARB_EN is asserted, the arbitration algorithm is switched from the LRU algorithm to the time-division algorithm. When the counter 35 counts the predetermined time (the counted value exceeds the predetermined value), the counted value becomes 0, and the arbitration switching signal ARB_EN is deasserted, and the arbitration algorithm is switched back to the original LRU algorithm. The arbitration switching signal ARB_EN is feed back to the OR circuit 34. On the other hand, the arbitration switching signal ARB_EN is input to a decoder 36, and the decoder 36 outputs the selection signals ARB_SEL1 and ARB_SEL2.
  • FIG. 7 is a diagram illustrating a structure of the queue arbiter circuit 28-1. The port arbiter circuit 28-2 may be realized by a circuit having the same structure as the queue arbiter circuit 28-1, and thus, the illustration and description of the structure of the port arbiter circuit 28-2 will be omitted. When using the arbiter circuit illustrated in FIG. 7 as the port arbiter circuit 28-2, the selection signal ARB_SEL2 is of course input in place of the selection signal ARB_SEL1.
  • The queue arbiter circuit 28-1 includes a LRU algorithm part 281, selectors 282 and 283, AND circuits 284 and 285, an OR circuit 286, and a counter 287 that are connected as illustrated in FIG. 7. A signal in accordance with the LRU algorithm part 281 and the selection signal ARB_SEL1 are input to the selector 282, and an output of the selector 282 is selected by the arbitration switching signal ARB_EN. For example, the queue AQ at the input port AI is input to the selector 283, and an output of the selector 283 is selected by the output signal of the selector 283. Each of the AND circuits 284 and 285 is input with the output signal of the selector 283 and an output port busy signal. A routing request (path control request) is output from the AND circuit 284.
  • A stall signal is output from the AND circuit 285 and is input to the OR circuit 286 and to a count-up terminal up of the counter 287. An output signal of the counter 287 is input to the LRU algorithm part 281 and the OR circuit 286, and is input in this case to the arbitration switching circuit 30 as a retry instruction with respect to the input port AI. Because an output signal of the OR circuit 286 is input to a clear terminal clr of the counter 287, the counter 287 is cleared when the counted value of the counter 287 reaches a predetermined value. The counter 287 may be formed by a 4-bit counter, for example.
  • Accordingly, the queue arbiter circuit 28-1 normally performs the arbitration according to the LRU algorithm of the LRU algorithm part 281, but switches to the arbitration according to the time-division algorithm when the arbitration switching signal ARB_EN is asserted. When the arbitration switching signal ARB_EN is deasserted after a predetermined time, the arbitration algorithm is returned back to the LRU algorithm from the time-division algorithm.
  • Next, a description will be given of how the livelock may be prevented in this embodiment, by referring to FIGS. 8 through 10. The crossbar switch 14 physically includes a plurality of ports, and a plurality of channels (that is, virtual channels) are logically included in each port, but at an arbitrary point in time, only one channel is selectable at one port.
  • FIG. 8 is a diagram illustrating an example of a state of the queue AQ at the input port AI. FIG. 9 is a diagram illustrating an arbitration result for each clock based on the LRU algorithm, and FIG. 10 is a diagram illustrating an arbitration result for each clock based on the time-division algorithm. FIGS. 9 and 10 illustrate an example in which the output port busy signal of the output port DO, the existence of the retry instruction, and the input ports AI, BI, CI are indicated together with the arbitration result. The busy state of the output port is indicated when the output port busy signal is "1". In addition, the existence of the retry instruction is indicated when the existence of the retry instruction is "1". Further, A1x and A2x indicate packets of the channels C0 and C1 at the input port AI, B1x and B2x indicate packets of the virtual channels C0 and C1 at the input port BI, and C1x and C2x indicate packets of the channels C0 and C1 at the input port CI.
  • For the sake of convenience, it is assumed that the packets A10, A11 and A12 of the channel C0 and the packets A20, A21 and A22 of the channel C1, intended for the output port DO, are all in the queue AQ at the input port AI and retained in the buffer within the crossbar switch 14. FIG. 8 only illustrates the queue AQ at the input port AI, however, the arbitration described hereunder is of course similarly performed with respect to the input ports BI and CI.
  • The arbitration from each of the input ports AI, BI and CI to the output port DO is normally performed according to the LRU algorithm, as illustrated in FIG. 9. FIG. 9 illustrates the arbitration in units of cycles, and in a cycle indicating that the output port busy signal is "1" and the output port DO is busy, the arbitration may not be performed because the resource of the output port DO is busy. In addition, in the cycle indicating that the existence of the retry instruction is "1" and the retry instruction exists, the routing request of each of the input ports AI, BI and CI is once cancelled before the arbitration is performed again. Accordingly, the cancelled packet has a lowest priority according to the LRU algorithm. In the example illustrated in FIG. 9, the resource of the output port DO is steadily busy, and the arbitration is only possible one packet at a time at predetermined intervals. In such a case, when viewed from the input port AI, only the packet A1x of the channel C0 in the queue AQ illustrated in FIG. 8 is processed, and the packet A2x of the channel C1 in the queue AQ remains retained in the buffer within the crossbar switch 14. Furthermore, when a subsequent packet is additionally input or inserted into the queue of the packet A1x of the channel C0, the livelock occurs with respect to the queue of the packet A2x of the channel C1.
  • On the other hand, when the time-division arbitration algorithm is used as illustrated in FIG. 10 in a situation identical to that described above for FIG. 9, the arbitration is fixedly performed for the same channel and the same input port at predetermined intervals, and such an arbitration is equally allocated to all channels of all input ports. For this reason, it may be seen that the livelock does not occur for some of the channels.
  • Although the performance of the parallel computer 1 deteriorates if the arbitration is simply performed using the time-division algorithm, the performance deterioration of the parallel computer 1 may be suppressed and the livelock may be prevented, by switching the arbitration algorithm from the LRU algorithm to the time-division algorithm only when there is a possibility of livelock.
  • According to the first embodiment described above, the switching of the arbitration algorithm is triggered when the frequency of the retry instruction reaches the threshold value. However, the trigger for the switching of the arbitration algorithm may be other than the frequency of the retry instruction.
  • FIG. 11 is a diagram illustrating an example of an arbiter circuit in a second embodiment. It is assumed for the sake of convenience that an arbiter circuit 28A illustrated in FIG. 11 is provided within the crossbar switch 14 of the parallel computer 1 illustrated in FIG. 1, for example. In FIG. 11, those parts that are the same as those corresponding parts in FIG. 5 are designated by the same reference numerals, and a description thereof will be omitted.
  • This embodiment differs from the first embodiment described above, in that lock information for each of the input ports AI, BI, CI and DI is supplied from a corresponding queue arbiter circuit 28-1A to an arbitration switching circuit 30A.
  • FIG. 12 is a diagram illustrating a structure of the queue arbiter circuit 28-1A. In FIG. 12, those parts that are the same as those corresponding parts in FIG. 7 are designated by the same reference numerals, and a description thereof will be omitted. In FIG. 12, the output signal of the selector 282, the output port busy signal, and the packet A1x (routing request) of the channel C0 at the input port AI are input to each of AND circuits 301 and 302. The output signal of the selector 282, the output port busy signal, and the packet A2x (routing request) of the channel C1 at the input port AI are input to each of AND circuits 311 and 312.
  • In FIG. 12, a high-level output signal of the AND circuit 301 indicates that the arbitration is not performed with respect to the packet A1x of the channel C0 at the input port AI, and the output signal of the AND circuit 301 is input to a count-up terminal up of a counter 303. A high-level output signal of the AND circuit 302 indicates that the arbitration is performed with respect to the packet A1x of the channel C0 at the input port AI, and the output signal of the AND circuit 302 is input to a clear terminal clr of the counter 303 via an OR circuit 304. The counter 303 may be formed by a 4-bit counter, for example. An output signal of the counter 303 is output to the arbitration switching circuit 30A as a port A1 lock signal to lock the packet A1x of the channel C0 at the input port AI, and is also fed back to the OR circuit 304. When a counted value of the counter 303 reaches a predetermined value, the output signal (port A1 lock signal) of the counter 303 indicates that the packet A1x of the channel C0 at the input port AI is in a deadlock state.
  • A high-level output signal of the AND circuit 311 indicates that the arbitration is not performed with respect to the packet A2x of the channel C1 at the input port AI, and the output signal of the AND circuit 311 is input to a count-up terminal up of a counter 313. A high-level output signal of the AND circuit 312 indicates that the arbitration is performed with respect to the packet A2x of the channel C1 at the input port AI, and the output signal of the AND circuit 312 is input to a clear terminal clr of the counter 313 via an OR circuit 314. The counter 313 may be formed by a 4-bit counter, for example. An output signal of the counter 313 is output to the arbitration switching circuit 30A as a port A2 lock signal to lock the packet A2x of the channel C1 at the input port AI, and is also fed back to the OR circuit 314. When a counted value of the counter 313 reaches a predetermined value, the output signal (port A2 lock signal) of the counter 313 indicates that the packet A2x of the channel C1 at the input port AI is in a deadlock state.
  • FIG. 13 is a diagram illustrating a structure of the arbitration switching circuit 30A. In FIG. 13, those parts that are the same as those corresponding parts in FIG. 5 are designated by the same reference numerals, and a description thereof will be omitted.
  • The arbitration switching circuit 30A includes OR circuits 330 and 34, the counter 35 to measure a predetermined time, and the decoder 36 that are connected as illustrated in FIG. 13. The lock information detected in each queue arbiter circuit 28-1A, that is, the port A1 lock signal and the port A2 lock signal with respect to the input port AI, a port B1 lock signal and a port B2 lock signal with respect to the input port BI, a port C1 lock signal and a port C2 lock signal with respect to the input port CI, and a port D1 lock signal and a port D2 lock signal with respect to the input port DI, are input to the OR circuit 330. An output signal of the OR circuit 330 is supplied to one input of the OR circuit 34.
  • When the lock information from each of the input ports AI, BI, CI and DI is input, the arbitration switching signal ARB_EN is asserted while the counter 35 counts the predetermined time (that is, until the counted value exceeds the predetermined value). During the time in which the arbitration switching signal ARB_EN is asserted, the arbitration algorithm is switched from the LRU algorithm to the time-division algorithm. When the counter 35 counts the predetermined time (the counted value exceeds the predetermined value), the counted value becomes 0 and the arbitration switching signal ARB_EN is deasserted, to thereby return the arbitration algorithm back to the original LRU algorithm.
  • Therefore, according to this embodiment, the arbitration algorithm may be switched using the deadlock in at least one of the input ports as the trigger.
  • FIG. 14 is a diagram illustrating an example of an arbiter circuit in a third embodiment. It is assumed that an arbiter circuit 28B illustrated in FIG. 14 is provided within the crossbar switch 14 of the parallel computer 1 illustrated in FIG. 1, for example. In FIG. 14, those parts that are the same as those corresponding parts in FIG. 5 are designated by the same reference numerals, and a description thereof will be omitted.
  • This embodiment differs from the first embodiment described above in that, empty signals EMPTY_A, EMPTY_B, EMPTY_C and EMPTY_D respectively indicating that the queues AQ, BQ, CQ and DQ at each of the input ports AI, BI, CI and DI are empty are supplied from the buffer to an arbitration switching circuit 30B. In addition, during a time in which the arbitration switching circuit 30B switches the arbitration algorithm from the LRU algorithm to the time-division algorithm, an inhibit signal INH that inhibits input or insertion of the packet is supplied to each of the input ports AI, BI, CI and DI.
  • FIG. 15 is a diagram illustrating a structure of the arbitration switching circuit 30B. In FIG. 15, those parts that are the same as those corresponding parts in FIG. 6 are designated by the same reference numerals, and a description thereof will be omitted.
  • In FIG. 15, the empty signals EMPTY_A, EMPTY_B, EMPTY_C and EMPTY_D respectively indicating that the queues AQ, BQ, CQ and DQ at each of the input ports AI, BI, CI and DI are empty are input from the buffer to an AND circuit 401. An output signal of the AND circuit 401 is input to the clear terminal clr of the counter 35. Hence, when all of the empty signals EMPTY_A, EMPTY_B, EMPTY_C and EMPTY_D from the queues AQ, BQ, CQ and DQ are asserted, the counted value of the counter 35 is cleared to 0 even before the measurement of the predetermined time ends. In addition, the arbitration switching signal ARB_EN output from the counter 35 is input to each of the input ports AI, BI, CI and DI as the inhibit signal INH via a buffer 402.
  • Hence, during a time in which the arbitration switching signal ARB_EN is asserted, the input or insertion of the packet to each of the input ports AI, BI, CI and DI is inhibited because the asserted inhibit signal INH is supplied to each of the input ports AI, BI, CI and DI. On the other hand, although the arbitration algorithm is switched from the LRU algorithm to the time-division algorithm during the time in which the arbitration switching signal ARB_EN is asserted, the counted value of the counter 35 is cleared to 0 at the time when all of the queues AQ, BQ, CQ and DQ at each of the input ports AI, BI, CI and DI become empty and all of the empty signals EMPTY_A, EMPTY_B, EMPTY_C and EMPTY_D are asserted. When the counted value of the counter 35 is cleared, the arbitration switching signal ARB_EN is deasserted, and thus, the arbitration algorithm is switched back to the original LRU algorithm.
  • Therefore, in this embodiment, the switching of the arbitration algorithm is triggered when the frequency of the retry instruction reaches the threshold value, and the switching of the arbitration algorithm back to the original arbitration algorithm is triggered when each of the queues AQ, BQ, CQ and DQ becomes empty and no more packet exists therein.
  • INDUSTRIAL APPLICABILITY
  • The disclosed arbitration method, arbiter circuit, and apparatus provided with the arbiter circuit may be applied to various apparatuses requiring an arbitration process, such as a crossbar switch, a parallel computer provided with the crossbar switch, and the like.
  • The disclosed arbitration method, arbiter circuit, and apparatus provided with the arbiter circuit are described by way of embodiments, however, the present invention is not limited to these embodiments, and various variations and modifications may be made without departing from the scope of the present invention.

Claims (15)

  1. An arbitration method comprising:
    a first arbitration process to perform a path control to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary output port amongst plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm;
    a second arbitration process to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and
    a switching process to switch the arbitrary arbitration algorithm used in the first and second arbitration processes to the time-division algorithm for a predetermined time in response to a trigger.
  2. The arbitration method as claimed in claim 1, wherein a retry instruction to once cancel a path control request from an arbitrary input port is output with respect to the arbitrary input port when the path control request from the arbitrary input port is not accepted in the second arbitration process for a predetermined time, in order to perform the arbitration again in the first arbitration process, and the trigger is generated when a frequency of the retry instruction reaches a threshold value.
  3. The arbitration method as claimed in claim 2, wherein the arbitration algorithm is returned to the arbitrary arbitration algorithm in response to another trigger that is generated when each queue at the plural input ports becomes empty and no more data exists therein.
  4. The arbitration method as claimed in claim 1, wherein the trigger is generated when a deadlock is detected in at least one of the plural input ports in the first arbitration process.
  5. The arbitration method as claimed in any of claims 1 to 4, wherein the arbitrary arbitration algorithm is selected from a group consisting of LRU (Least Recently Used) algorithm, round robin algorithm, and fixed priority algorithm.
  6. An arbiter circuit comprising:
    a first arbiter circuit configured to perform a path control to transfer data from physically plural input ports logically having plural virtual channels to an arbitrary output port amongst plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm;
    a second arbiter circuit configured to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and
    an arbitration switching circuit configured to switch the arbitrary arbitration algorithm used in the first and second arbiter circuits to the time-division algorithm for a predetermined time in response to a trigger.
  7. The arbiter circuit as claimed in claim 6,
    wherein a retry instruction to once cancel a path control request from an arbitrary input port is output with respect to the arbitrary input port when the path control request from the arbitrary input port is not accepted in the second arbiter circuit for a predetermined time, in order to perform the arbitration again in the first arbiter circuit, and
    wherein the arbitration switching circuit uses the trigger that is generated when a frequency of the retry instruction reaches a threshold value.
  8. The arbiter circuit as claimed in claim 7, wherein the arbitration switching circuit returns the arbitration algorithm to the arbitrary arbitration algorithm in response to another trigger that is generated when each queue at the plural input ports becomes empty and no more data exists therein.
  9. The arbiter circuit as claimed in claim 6, wherein the arbitration switching circuit uses the trigger that is generated when a deadlock is detected in at least one of the plural input ports in the first arbiter circuit.
  10. The arbiter circuit as claimed in any of claims 6 to 9, wherein the arbitrary arbitration algorithm is selected from a group consisting of LRU (Least Recently Used) algorithm, round robin algorithm, and fixed priority algorithm.
  11. An apparatus to couple a system having a plurality of processors and memories to an input output system having a plurality of input output interfaces, comprising:
    physically plural input ports logically having plural virtual channels;
    plural output ports;
    a first arbiter circuit configured to perform a path control to transfer data from the plural input ports to an arbitrary output port amongst the plural output ports, wherein only one channel is selectable at one input port at an arbitrary point in time, by performing an arbitration among the channels of each of the plural input ports according to an arbitrary arbitration algorithm other than a time-division algorithm;
    a second arbiter circuit configured to perform an arbitration among the plural input ports according to the arbitrary arbitration algorithm; and
    an arbitration switching circuit configured to switch the arbitrary arbitration algorithm used in the first and second arbiter circuits to the time-division algorithm for a predetermined time in response to a trigger.
  12. The apparatus as claimed in claim 11,
    wherein a retry instruction to once cancel a path control request from an arbitrary input port is output with respect to the arbitrary input port when the path control request from the arbitrary input port is not accepted in the second arbiter circuit for a predetermined time, in order to perform the arbitration again in the first arbiter circuit, and
    wherein the arbitration switching circuit uses the trigger that is generated when a frequency of the retry instruction reaches a threshold value.
  13. The apparatus as claimed in claim 12, wherein the arbitration switching circuit returns the arbitration algorithm to the arbitrary arbitration algorithm in response to another trigger that is generated when each queue at the plural input ports becomes empty and no more data exists therein.
  14. The apparatus as claimed in claim 11, wherein the arbitration switching circuit uses the trigger that is generated when a deadlock is detected in at least one of the plural input ports in the first arbiter circuit.
  15. The apparatus as claimed in any of claims 11 to 14, wherein the arbitrary arbitration algorithm is selected from a group consisting of LRU (Least Recently Used) algorithm, round robin algorithm, and fixed priority algorithm.
EP09842617A 2009-03-31 2009-03-31 Arbitration method, arbitration circuit, and device using the arbitration circuit Withdrawn EP2416254A4 (en)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2009/056671 WO2010113262A1 (en) 2009-03-31 2009-03-31 Arbitration method, arbitration circuit, and device using the arbitration circuit

Publications (2)

Publication Number Publication Date
EP2416254A1 true EP2416254A1 (en) 2012-02-08
EP2416254A4 EP2416254A4 (en) 2012-10-17

Family

ID=42827595

Family Applications (1)

Application Number Title Priority Date Filing Date
EP09842617A Withdrawn EP2416254A4 (en) 2009-03-31 2009-03-31 Arbitration method, arbitration circuit, and device using the arbitration circuit

Country Status (4)

Country Link
US (1) US8837505B2 (en)
EP (1) EP2416254A4 (en)
JP (1) JP5304888B2 (en)
WO (1) WO2010113262A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105099956A (en) * 2015-06-25 2015-11-25 华为技术有限公司 Exchange network system and data exchange method

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5471627B2 (en) * 2010-03-09 2014-04-16 富士通株式会社 Network device, edge router and packet communication system
US9678906B2 (en) * 2014-03-26 2017-06-13 International Business Machines Corporation Oldest link first arbitration between links grouped as single arbitration elements
JP6350098B2 (en) * 2014-08-11 2018-07-04 富士通株式会社 Arithmetic processing device, information processing device, and control method for information processing device
US10877836B2 (en) * 2018-08-29 2020-12-29 Arm Limited Method and apparatus for coherent interconnect recovery with protocol layer re-transmission

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030227901A1 (en) * 2002-06-10 2003-12-11 Kodialam Muralidharan S. Scheduling of guaranteed-bandwidth low-jitter traffic in input-buffered switches
US20080034140A1 (en) * 2004-06-16 2008-02-07 Koji Kai Bus Arbitrating Device and Bus Arbitrating Method

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH10254843A (en) * 1997-03-06 1998-09-25 Hitachi Ltd Crossbar switch, parallel computer with the crossbar switch and broadcasting communication method
JP2001022711A (en) 1999-07-08 2001-01-26 Hitachi Ltd Method for controlling crossbar switch and parallel computer system
US20020146022A1 (en) * 2000-05-31 2002-10-10 Van Doren Stephen R. Credit-based flow control technique in a modular multiprocessor system
JP3747020B2 (en) * 2002-09-09 2006-02-22 エヌイーシーコンピュータテクノ株式会社 Crossbar mediation system
JP4907166B2 (en) * 2005-01-12 2012-03-28 パナソニック株式会社 Resource management device
US7395361B2 (en) * 2005-08-19 2008-07-01 Qualcomm Incorporated Apparatus and methods for weighted bus arbitration among a plurality of master devices based on transfer direction and/or consumed bandwidth

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030227901A1 (en) * 2002-06-10 2003-12-11 Kodialam Muralidharan S. Scheduling of guaranteed-bandwidth low-jitter traffic in input-buffered switches
US20080034140A1 (en) * 2004-06-16 2008-02-07 Koji Kai Bus Arbitrating Device and Bus Arbitrating Method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of WO2010113262A1 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105099956A (en) * 2015-06-25 2015-11-25 华为技术有限公司 Exchange network system and data exchange method
CN105099956B (en) * 2015-06-25 2018-08-14 华为技术有限公司 Switching network system and method for interchanging data
US10505867B2 (en) 2015-06-25 2019-12-10 Huawei Technologies Co., Ltd. Switch fabric system and data switching method

Also Published As

Publication number Publication date
JP5304888B2 (en) 2013-10-02
WO2010113262A1 (en) 2010-10-07
JPWO2010113262A1 (en) 2012-10-04
US20120002677A1 (en) 2012-01-05
EP2416254A4 (en) 2012-10-17
US8837505B2 (en) 2014-09-16

Similar Documents

Publication Publication Date Title
US7433363B2 (en) Low latency switch architecture for high-performance packet-switched networks
EP2416254A1 (en) Arbitration method, arbitration circuit, and device using the arbitration circuit
US6246256B1 (en) Quantized queue length arbiter
US8234428B2 (en) Arbitration device that arbitrates conflicts caused in data transfers
US9262356B2 (en) Arbiter device and arbitration method
US8051234B2 (en) Multiprocessor system
US10614010B2 (en) Handling queued interrupts in a data processing system based on a saturate value
EP0549248A2 (en) An arbitration system and method
US6700899B1 (en) Bit slice arbiter
US20040103232A1 (en) Arbitration unit
US20140304450A1 (en) Switching device, packet control method, and data communication system
US7002981B2 (en) Method and arbitration unit for digital switch
US9996489B2 (en) Memory aggregation device
EP2290552A1 (en) Data transfer device, information processing device, and control method
JP4669174B2 (en) Tunable broadcast / point-to-point packet arbitration
US20030152082A9 (en) Distribution of weightings between port control system and switch cards of a packet switching device
EP2854042A1 (en) Information processing apparatus, data transfer apparatus, and data transfer method
CN104978289A (en) Service Request Interrupt Router With Shared Arbitration Unit
JP4679601B2 (en) Packet control circuit, packet processing apparatus, and packet processing method
JPH0528103A (en) Bus multiplex system
JP2007026284A (en) Data processor
JP2000201161A (en) Arbitration control circuit
JP2001337841A (en) Task control unit
JP2001285347A (en) Device for packet switching and method for statistics information processing
JP2008306609A (en) Switching arbitration method, switching arbitration apparatus, switching arbitration program, and program recording medium

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 20111024

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO SE SI SK TR

DAX Request for extension of the european patent (deleted)
A4 Supplementary search report drawn up and despatched

Effective date: 20120919

RIC1 Information provided on ipc code assigned before grant

Ipc: H04L 12/56 20060101AFI20120912BHEP

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION HAS BEEN WITHDRAWN

18W Application withdrawn

Effective date: 20141106